Feature extraction and clustering of EEG epileptic spikes

This paper treats algorithms for feature extraction and clustering of multichannel EEG transients occurring in epilepsy, so called spikes. Hermite functions with a variable width parameter is used as features. We study nonlinear optimization of a series expansion for multichannel spikes. For the clustering problem, the nearest mean (NM) algorithm, generalized to matrix features, is used. The number of classes is assumed to be known a priori. The series expansion gives good signal description while reducing information. A simulation to estimate the space resolution capability of the algorithms indicates that perfect clustering requires approximately one head radius distance between the dipoles, which each generate one cluster. The NM algorithm was used to cluster two sets of clinically recorded spikes, and the clustering was compared to the manual clustering obtained by a neurophysiologist. For both spike sets evaluated, the clusters obtained by the algorithms had high accordance with the result of the neurophysiologist.